To address the limitations of weak interpretability and poor generalization in existing taxi-out time prediction models, a novel prediction model for departing flights based on Stacking ensemble learning and Shapley additive explanations (SHAP) was proposed. Firstly, the taxi-out time was decomposed into unimpeded and dynamic components, with their correlations to influencing factors analyzed separately. Subsequently, a Stacking-based prediction model was constructed, and a comparative evaluation was conducted between holistic and phased prediction approaches. Finally, SHAP analysis was implemented to quantify the importance of feature variables in the Stacking model, and the rationality of the model was validated using actual operational data from Shenzhen Bao’an International Airport. The results indicate that unimpeded taxi-out time is mainly influenced by the configuration of the airport, while dynamic taxi-out time is primarily affected by surface traffic flow. Although the phased prediction approach exhibits marginally inferior performance compared to holistic prediction, it offers stronger interpretability. Mean absolute percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE) values of 10.6%, 99.7 s, and 140.5 s, respectively, are achieved by the Stacking model, with prediction accuracy rates of 41.0%, 86.3%, and 96.5% for intervals of ±60, ±180, and ±300 s, demonstrating superior prediction accuracy and generalization capability compared to existing studies. Feature dimensionality is effectively reduced by a dual feature selection mechanism combining SHAP analysis and correlation analysis, while high prediction accuracy is maintained.